Home | Search | Help  
Home Page Università di Genova

Seminar Details

Date 4-7-2011
Time 11:15
Room/Location DISI-Sala Conferenze III piano
Title Random Forests: A Swiss Army Knife Classifier
Speaker Leif Peterson, Ph.D.
Affiliation Associate Member The Methodist Hospital Research Institute Director, Center for Biostatistics Associ
Link http://www.methodisthealth.com/tmhri.cfm?id=37614
Abstract Random Forest (RF) is a stable classifier which is not biased toward the amount of training data used, and offers little chance of overfitting the data.Generalization error is, for the most part, lower than other classifiers.RF classification accuracy values tend to be lower than other classifiers because unlike other classifiers RF randomly samples objects and features used for each decision tree in the forest.This presentation will introduce class discovery with unsupervised classification via RF, class prediction with supervised RF classification, feature importance scores, and object outlier determination.RF offers many advantages over other approaches including reduced generalization error, class discovery through unsupervised cluster prediction, class prediction via supervised learning, simultaneous interaction effects of multiple features, better handling of missing data, and rapid learning times. One downturn of RF is that it is not necessary to perform cross validation, e.g., 10-fold or leave-one-out, due to the advantage of using bootstrapping.Bootstrapping has been shown to result in low variance or less variation across data sets, and potentially greater levels of bias due to its dependence on the proportion of data used for training
Back to Seminars